> ## Documentation Index > Fetch the complete documentation index at: https://docs.priorlabs.ai/llms.txt > Use this file to discover all available pages before exploring further. > ## Agent Instructions > TabPFN-3 and later versions support up to 1,000,000 rows, subject to feature count and checkpoint/API limits. > For large datasets, use per-estimator subsampling, e.g. inference_config={"SUBSAMPLE_SAMPLES": 100_000}, in TabPFNClassifier or TabPFNRegressor from tabpfn or tabpfn_client. > Pass raw pandas DataFrames to fit and predict. Categorical strings/categories, raw text, and missing feature values are handled automatically; no manual integer/one-hot encoding, imputation, scaling, or outlier removal is needed. # TabPFN vs XGBoost > Compare TabPFN and XGBoost on accuracy, tuning effort, training time, prediction speed, hardware, data limits, and production trade-offs. XGBoost uses [gradient-boosted decision trees](https://en.wikipedia.org/wiki/Gradient_boosting), training a model specifically for each dataset. [TabPFN](https://en.wikipedia.org/wiki/TabPFN) is a pretrained transformer that makes predictions on new datasets without traditional task-specific training. ## TabPFN vs XGBoost at a glance | | TabPFN | XGBoost | | - | - | - | | **Approach** | Pretrained transformer (TabPFN) | Gradient-boosted trees | | **Training** | No training required | Trains on each dataset | | **Tuning** | Minimal | Often needed | | **Prediction speed** | Under 0.2ms on 1M rows | Typically fast | | **Best suited for** | Best predictive performance & fast deployment | CPU-efficient deployment | ## Which should you choose? ### Choose TabPFN if... ### Choose XGBoost if... ## Predictive Performance Comparison TabPFN and XGBoost take different approaches to learning from tabular data, but how does that translate into predictive performance? To compare them, we look at results from TabArena, a benchmark evaluating tabular models across classification and regression tasks. Its Elo ratings summarize relative predictive performance across benchmark tasks. Higher Elo indicates stronger relative performance. The results below compare TabPFN-3.5 and TabPFN-3.5-Fast against both default and tuned-and-ensembled XGBoost configurations. Classification Regression
ModelElo95% Confidence Interval
TabPFN-3.5, default1861+107/-70
TabPFN-3.5-Fast, default1795+97/-70
XGBoost, tuned + ensembledTuned-and-ensembled XGBoost includes additional optimization and post-hoc ensembling effort.1371+40/-49
XGBoost, defaultDefault configurations are evaluated without extensive hyperparameter optimization.1227+39/-44
Higher Elo is better. Configurations are listed separately where the board reports default and tuned results. *Data as of 2026-10-28. [Source CSV](https://huggingface.co/spaces/TabArena/leaderboard/resolve/68988c6e5b31e2ec9a2d193c555dd41b8cb01fd3/data/entrants_open/imputation_no/splits_all/tasks_classification/datasets_all/website_leaderboard.csv). Revision: `68988c6e5b31e2ec9a2d193c555dd41b8cb01fd3`.* On TabArena, TabPFN-3.5 outperforms both default and tuned-and-ensembled XGBoost in classification and regression rankings. Notably, TabPFN achieves these results without dataset-specific hyperparameter tuning, while the tuned XGBoost configuration includes additional optimization and ensembling. ## Training and prediction speed Predictive performance is only part of the comparison. The time required to prepare a model and generate predictions also matters. XGBoost trains decision trees on each dataset. TabPFN uses a pretrained model, avoiding traditional task-specific weight training. This means that achieving stronger performance with XGBoost often involves hyperparameter tuning (testing different model configurations) and ensembling (combining predictions from multiple trained models), both of which can significantly increase training time. Classification Regression
ModelFit/setup (s/1K)Predict (s/1K)Hardware
TabPFN-3.5, default1.850.518GPU
TabPFN-3.5-Fast, default0.720.168GPU
XGBoost, tuned + ensembledTuned-and-ensembled XGBoost includes additional optimization and post-hoc ensembling effort.685.871.454CPU
XGBoost, defaultDefault configurations are evaluated without extensive hyperparameter optimization.1.770.117CPU
TabArena median times per 1,000 rows. Fit/setup time for TabPFN is per-dataset setup and context processing rather than weight training. *Data as of 2026-10-28. [Source CSV](https://huggingface.co/spaces/TabArena/leaderboard/resolve/68988c6e5b31e2ec9a2d193c555dd41b8cb01fd3/data/entrants_open/imputation_no/splits_all/tasks_classification/datasets_all/website_leaderboard.csv). Revision: `68988c6e5b31e2ec9a2d193c555dd41b8cb01fd3`.* **Note:** These timings reflect the hardware used in the benchmark: TabPFN runs on GPU, while XGBoost runs on CPU. They illustrate the benchmark configurations rather than a hardware-controlled speed comparison. ## Native data types TabPFN-3.5 supports common tabular feature types and includes built-in preprocessing for text and datetime columns, reducing the need for manual feature engineering: | Data type | TabPFN-3.5 | XGBoost | | :- | :- | :- | | Numerical | Native support | Native support | | Categorical | Native support | Native support (`enable_categorical=True`) | | Missing values | Native support | Native support | | Text | Built-in preprocessing | Requires feature engineering | | Datetime | Built-in preprocessing | Requires feature engineering | ## Dataset characteristics Beyond raw feature types, dataset characteristics such as class imbalance, temporal structure, and distribution shift can affect model performance and the amount of additional configuration required. The table below compares how TabPFN and XGBoost approach these challenges. | Dataset characteristic | TabPFN-3.5 | XGBoost | | :- | :- | :- | | Small datasets | Strong out of the box | Prone to overfitting without careful regularization and tuning | | Imbalanced classes | Robust by default, with optional [metric tuning](https://docs.priorlabs.ai/capabilities/metric-tuning) (`balance_probabilities`, `eval_metric`) | Requires manual handling (e.g. `scale_pos_weight`, resampling) | | Time series / forecasting | Supported via [TabPFN-TS](https://docs.priorlabs.ai/capabilities/forecasting.md) and [related-series forecasting](https://docs.priorlabs.ai/cookbook/forecast_hierarchical_series.md) | Requires feature engineering (lags, rolling windows, calendar features) | | Grouped data | Strong results on grouped splits in BeyondArena; see [grouped data comparison](https://docs.priorlabs.ai/cookbook/grouped_data_model_comparison.md) | Prone to overfitting groups without careful group-aware splits and tuning | | Distribution shift | Adapts in-context | Requires retraining and re-tuning on new data | > [!NOTE] > > Performance depends on the dataset, evaluation split, and metric. [BeyondArena](https://huggingface.co/datasets/TabArena/BeyondArena) evaluates tabular models across different dataset characteristics, including temporal and grouped splits. > **[See the difference in action](https://docs.priorlabs.ai/cookbook/tabpfn_vs_xgboost.md)** > > Compare TabPFN and XGBoost on a real dataset from OpenML. ## Frequently asked questions > **Is TabPFN better than XGBoost?** > > On TabArena, TabPFN-3.5 outperforms default and tuned-and-ensembled XGBoost in both classification and regression. However, performance depends on your dataset, so benchmark both models on your own data before choosing. > **Does TabPFN require training?** > > TabPFN is pretrained and does not require traditional task-specific weight training. Its .fit() method prepares the dataset for inference rather than training model weights, unlike XGBoost, which trains decision trees on each dataset. > **Can TabPFN replace XGBoost?** > > For many supervised tabular prediction tasks, TabPFN can be an alternative to XGBoost, particularly when you want strong performance with minimal tuning. However, XGBoost may be preferable for very large datasets, CPU-only environments, or applications requiring specific high-throughput inference. > **Is TabPFN faster than XGBoost?** > > It depends on what you’re measuring. TabPFN avoids traditional model training, which can reduce time to a strong model. XGBoost typically offers efficient prediction after training, particularly on CPUs. Compare end-to-end runtime and prediction latency on your hardware. > **How can I compare TabPFN and XGBoost on my own dataset?** > > Use identical train/test splits, evaluation metrics, and comparable computational budgets. Evaluate both models using validation or cross-validation, then report final performance on an untouched test set. Follow the [benchmarking guide](https://docs.priorlabs.ai/benchmarking.md) or the [TabPFN vs XGBoost cookbook](https://docs.priorlabs.ai/cookbook/tabpfn_vs_xgboost.md). This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.